Leaker discusses Nvidia Blackwell memory configurations — suggests GB202 with 512-bit, GB203 with 256-bit, and GB205 with 192-bit

GeForce RTX 4090
(Image credit: Nvidia)

GPU detective Kopite7kimi on X suggests that Nvidia's upcoming RTX 50-series "Blackwell" GB203 and new GB205 dies will feature the same memory bus widths as Nvidia's outgoing RTX 40-series AD103 and AD104 GPU dies — those are found in the likes of the RTX 4080 Super and the RTX 4070 Super, some of the best graphics cards.

Kopite7kimi's latest post appears to be a confirmation of what he said a few days ago, regarding Blackwell's memory bus sizes. Instead of widening or shortening the bus-width of its RTX 50-series GPUs, Kopite believes that Nvidia is content with the memory bus-width configurations it's using on its RTX 40-series Ada Lovelace GPUs and will translate those to its new GPUs based on the Blackwell architecture.

Kopite specifically states that at least two Blackwell's GPUs, GB203 and GB205, will share the same memory bus widths as their predecessors. GB203 will be the direct successor to AD103 used in the RTX 4080 series and the RTX 4070 Ti Super featuring a 256-bit interface. GB205 is more of an oddball, since there have been no "205" class GPUs in the past, but it is rumored to be a brand new GPU model that will not directly succeed AD104. Nonetheless, Kopite believes it will sport the same memory configuration as AD104 found in the RTX 4070 series and RTX 4070 Ti sporting a 192-bit interface.

There's no mention of a GB204 die, however, which would logically fall between the GB203 and GB205 in terms of specs. Perhaps Nvidia will skip that potential GPU completely, just as it didn't have "AD105" or "GA105" GPUs for the past two generations. All of these are rumors and speculations, of course, so take everything with a healthy dose of skepticism.

Swipe to scroll horizontally
Rumored Blackwell GPU memory specifications
Blackwell GPUWidthTypeCapacity (16 Gb | 24 Gb)Ada GPUWidthTypeCapacityAmpere GPUWidthTypeCapacity
GB202384-bitGDDR724 GB | 36 GBAD102384-bitGDDR6X24 GBGA102384-bitGDDR6X24 GB
GB203256-bitGDDR716 GB | 24 GBAD103256-bitGDDR6X16 GBGA103256-bitGDDR6X16 GB
----AD104192-bitGDDR6X12 GBGA104256-bitGDDR6X16 GB
GB205192-bitGDDR712 GB | 18 GB--------
GB206128-bit??AD106128-bitGDDR616 GBGA106192-bitGDDR616 GB
GB207128-bit??AD107128-bitGDDR68-16 GBGA107128-bitGDDR68-16 GB

What does this mean in terms of real-world performance? Not much on its own. We believe Nvidia will be transitioning to the GDDR7 graphics memory standard with the RTX 50-series. GDDR7 is a brand-new standard that was formalized by JEDEC just a week ago, featuring twice the bandwidth potential of GDDR6X (let alone GDDR6).

GDDR7's immense bandwidth potential will allow Nvidia to reduce/reuse its outgoing memory interfaces on its next-generation GPUs while at the same time substantially boosting bandwidth. Combined with larger L2/L3 cache sizes, that's one reason memory interface widths have been shrinking in recent years on both Nvidia and AMD GPUs. Larger caches and newer memory technologies will enable Nvidia to reduce the bus width while increasing effective memory bandwidth.

It's also important to point out that GDDR7 opens the door to several other potential upgrades. While GDDR6 (and GDDR6X) topped out at a theoretical 32Gb capacity per chip (4GB), GDDR7 will support up to 64Gb (8GB) memory sizes. That's probably a ways off, and most GDDR6 solutions are still using 16Gb (2GB) devices.

Perhaps more importantly, GDDR7 also supports non-power-of-two capacities like 24Gb and 48Gb, and Samsung and Micron have already indicated their initial GDDR7 devices will top out at 16Gb/24Gb. The use of 24Gb (3GB) devices would mean the GB203 could support up to 24GB and GB205 could support up to 18GB, with GB206/207 supporting 12GB. That would go a long way towards alleviating concerns about VRAM capacity, even if the memory bus widths stay the same.

Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

  • usertests
    This story seems disjointed. 512-bit is mentioned in the headline and then nowhere else.

    A fun move might be 512-bit 5090 with 32 GB, then the next generation they dial it back to 384-bit, but with an increase to 36 GB since 24 Gb GDDR7 will be ready by then.

    The 512-bit rumor contradicts previous rumors of 384-bit and 24 GB. Nvidia could be putting out misleading info behind the scenes again to trick leakers.
    Reply
  • DougMcC
    The nearly same story a couple of days ago claiming 384it:
    https://www.tomshardware.com/pc-components/gpus/next-gen-nvidia-geforce-gaming-gpu-memory-spec-leaked-rtx-50-blackwell-series-gb20x-memory-configs-shared-by-leaker
    Reply
  • oofdragon
    DougMcC said:
    The nearly same story a couple of days ago claiming 384it:
    https://www.tomshardware.com/pc-components/gpus/next-gen-nvidia-geforce-gaming-gpu-memory-spec-leaked-rtx-50-blackwell-series-gb20x-memory-configs-shared-by-leaker
    Exactly.. they just care about the clicks
    Reply
  • thisisaname
    <blockquote> The use of 24Gb (3GB) devices would mean the GB203 could support up to 24GB and GB205 could support up to 18GB, with GB206/207 supporting 12GB. That would go a long way towards alleviating concerns about VRAM capacity, even if the memory bus widths stay the same.</blockquote>

    Or it could use 16Gb(2GB) then it would be 16GB, 12GB and 8GB and nothing much would have changed with regards to alleviating concerns around VRAM capacity.
    Reply
  • ilukey77
    personally i think i might go big big next gen love my AMD cards but if the 5090 is going to be a monster monster then i may try to swap out of the rat race go the 5090 and sit on it for 2 gens //
    providing im happy with the cable config im still worried about the current plug con fig and skeptical about the new one ..
    That said i burnt out ( super upset about it ) one of the plugs on my 7900xtx the other day using the pcie 6x2 plug so !!
    Reply